CONTEXT-BASED QUOTATION RECOMMENDATION - ANSEL MACLAUGHLIN1 , TAO CHEN2 , BURCU KARAGOL AYAN2, DAN ROTH3 - ANSEL ...
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Context-Based Quotation Recommendation Ansel MacLaughlin1∗ , Tao Chen2† , Burcu Karagol Ayan2 , Dan Roth3 1 Northeastern University, Boston, MA, USA 2 Google Research, Mountain View, CA, USA 3 University of Pennsylvania, Philadelphia, PA, USA ansel@ccs.neu.edu, {taochen, burcuka}@google.com, danroth@cis.upenn.edu Abstract Document Title: Return engagement to White House for Mi- ami Heat While composing a new document, anything from a news ar- Document Context: ... The president opened by recounting the ticle to an email or essay, authors often utilize direct quotes numerous accomplishments of the team last season, including from a variety of sources. Although an author may know what Miami’s team-record 66 wins. [Recommend Quote Here] point they would like to make, selecting an appropriate quote Source Paragraph: Last season, the Heat put together one for the specific context may be time-consuming and difficult. of the most dominating regular seasons ever by a defending We therefore propose a novel context-aware quote recom- champion. They won a team-record 66 games. At one point, mendation system which utilizes the content an author has al- they won 27 games straight- the second-longest winning streak ready written to generate a ranked list of quotable paragraphs ever, extraordinarily impressive- almost as impressive as the and spans of tokens from a given source document. Bulls’ 72-win season. Riley and I were reminiscing about those We approach quote recommendation as a variant of open- Knicks years. domain question answering and adapt the state-of-the-art BERT-based methods from open-QA to our task. We conduct Table 1: Given a title and context in a new document, mod- experiments on a collection of speech transcripts and asso- els aim to recommend both relevant source paragraphs and ciated news articles, evaluating models’ paragraph ranking quotable spans of text in them, such as in this example. The and span prediction performances. Our experiments confirm bold portion of paragraph was quoted by the author in the the strong performance of BERT-based methods on this task, next sentence of the document. which outperform bag-of-words and neural ranking baselines by more than 30% relative across all ranking metrics. Quali- tative analyses show the difficulty of the paragraph and span recommendation tasks and confirm the quotability of the best searching through them to select an appropriate quote may BERT model’s predictions, even if they are not the true se- be time-consuming and challenging. lected quotes from the original news articles. In this paper we propose a novel context-aware quote rec- ommendation (QR) task – recommending quotable passages Introduction of text from a source document based on the content an au- thor has already written. Previously proposed QR systems Machine-assisted writing tools have many potential uses, have either made predictions without using the textual con- including helping users compose documents more quickly text (Niculae et al. 2015) or only predicted from a fixed list (Hard et al. 2019), generate creative story ideas (Clark et al. of popular quotes, proverbs, and maxims (Lee et al. 2016; 2018), and improve their grammar (Rozovskaya and Roth Tan, Wan, and Xiao 2015, 2016; Ahn et al. 2016; Tan et al. 2014; Hoover, Lytvyn, and Shevchenko 2015). However, 2018). Our proposed task is more general and difficult, as writing a document, such as a news article, academic es- models must both identify quotable passages of arbitrary say, or email, consists not only of composing new text, but length in any given source document and recommend the of weaving in content from source documents by summa- most relevant ones to the author. Authors across a variety of rizing, paraphrasing, and quoting. Quotes are a particularly domains could benefit from such a system, including jour- important part of many documents, functioning to “change nalists quoting from an interview, speech or meeting tran- the pace of a story, add colour and character, illustrate bald script, students quoting from a chapter of a primary source facts, and introduce personal experience” (Cole 2008). How- for an essay, or social media posters writing about a recent ever, the source documents from which an author could se- blog post or news article. lect a quote may be long, consisting of hundreds of para- We approach the problem at two different granularities, graphs, and, as described in the findings of our pilot study, recommending both entire paragraphs in a source document ∗ Work completed while interning at Google Research. for an author to excerpt and specific spans of text in those † Corresponding author. paragraphs that are suitable for quoting (i.e. quotable). Ta- Copyright c 2021, Association for the Advancement of Artificial ble 1 shows an example recommendation – given the title Intelligence (www.aaai.org). All rights reserved. and a snippet of context from a new document, we aim to
recommend both a relevant and quotable source paragraph ment, and quote recommendation, recommending quotes to and specific subspan that an author can quote in the next sen- an author as they compose a new document. No earlier work tence of their document. The two granularities of recommen- has addressed both of these problems simultaneously. Prior dation offer different and complementary functionality to quotability identification research has focused only on iden- authors. Recommending relevant and quotable source para- tifying generally quotable content, regardless of the context graphs would allow authors to find suitable quotes faster, po- in which its used (Tan, Peng, and Smith 2018; Bendersky tentially searching over only the top ranked paragraphs for a and Smith 2012; Danescu-Niculescu-Mizil et al. 2012; Nic- span of tokens to excerpt. Further, since, as discussed in the ulae et al. 2015). On the other hand, previously proposed QR findings of our pilot study and error analysis, authors often models do not identify quotable content – they only make paraphrase surrounding content from the paragraph in the recommendations from a predefined list of quotable phrases introduction to and discussion of a quote, the paragraph’s (Tan, Wan, and Xiao 2015, 2016; Lee et al. 2016; Ahn et al. entire text beyond the specific span is still useful. Recom- 2016; Tan et al. 2018). mending spans of text, although more difficult than para- Unlike the proposed BERT models, previous quotability graphs, would save an author further time by allowing them identification systems have used linear classifiers with man- to easily scan through a ranked list of potential quotes. ually designed lexical and syntactic features to distinguish We cast our novel approach to QR as a variant of open- between quotable and un-quotable passages. Prior work has domain question answering (open-QA). In open-QA, for a studied applications in a variety of domains, including polit- given question, a model must retrieve relevant passages from ical speeches and debates (Tan, Peng, and Smith 2018; Nic- a source corpus and identify answer spans in those passages ulae et al. 2015), books (Bendersky and Smith 2012) and (Chen et al. 2017). Similarly, for our task, for a given con- movie scripts (Danescu-Niculescu-Mizil et al. 2012). text in a new document, a model must both identify relevant Both collaborative filtering and content-based approaches paragraphs in a source document and predict quotable spans have been used for QR systems. Unlike our work, which rec- of tokens in them. Following this intuition, we adapt state- ommends quotes to a specific context in a new document, of-the-art (SoTA) BERT-based open-QA methods for our Niculae et al. (2015) approach the problem as matrix com- task (Devlin et al. 2019). We fine-tune two separate BERT pletion and learn to predict which quotes a given news out- models, a paragraph ranker and a span predictor (Wang et al. let will report. They do not use textual features for predic- 2019), and combine their predicted scores to rank source tion and thus cannot generalize to new, previously unseen, paragraphs and spans by their predicted quotability. quotes. We conduct experiments on a dataset of White House More similar to our work are content-based QR systems speech and press conference transcripts and associated news that use textual content to recommend quotes for a given articles (Niculae et al. 2015). To first assess the difficulty of context in a new document (Tan, Wan, and Xiao 2015, 2016; the QR task, we recruit two journalists for a pilot study. We Lee et al. 2016; Ahn et al. 2016; Tan et al. 2018). However, find that identifying the “correct” (actually quoted) speech prior work approach the problem not as ranking paragraphs paragraphs and spans is hard, even for experts, and that, un- and predicting spans in a source document, but as ranking like QA, there may be multiple relevant and quotable para- a predefined list of popular quotes, proverbs and maxims. graphs and spans for a given context. Next, we adapt and Furthermore, unlike BERT, previous content-based systems apply SoTA BERT-based open-QA methods to the task. We neither deeply model the interactions between a context and find that the combined BERT model (Wang et al. 2019) per- quote nor identify the textual features in a quote indicative forms best, significantly outperforming bag-of-words and of its overall quotability. Finally, there are substantial mod- neural ranking baselines. Finally, in light of the findings of eling differences between our proposed system, which uses our pilot study, we perform a large-scale crowdsourced evau- only the textual content of both the context and paragraph lation of the quality of our best model’s top recommenda- to make a recommendation, and prior work. Tan, Wan, and tions, focusing on instances where the model misranks the Xiao (2015; 2016; 2018) train their models with various paragraph chosen by the original author. We find that human additional metadata features, such as topic and popularity raters judge the model’s top predicted (non ground truth) data crawled from the web. Lee et al. (2016) and Ahn et al. paragraph to be relevant and quotable 85% of the time, fur- (2016) approach recommendation as multi-class classifica- ther demonstrating the usefulness of our system. tion, using only the text of the context to predict a specific Our contributions include a novel approach to QR that quote class. These differences make previous systems in- generalizes to source documents not seen during training, compatible with our proposed task, where metadata may not an application of SoTA BERT-based open-QA methods that be available and models must rank an arbitrary number of combine paragraph and span-level models to capture com- source paragraphs not seen during training. plementary aspects of a passage’s quotability, and a qualita- tive evaluation methodology aimed at better dealing with the Citation Recommendation Our proposed QR system has fact that quote selection is subjective. similarities to content-based citation recommendation (CR) systems. These systems recommend citations for a new doc- Related Work ument based on its textual similarity to the candidate cita- Quotability Identification & Quote Recommendation tions. Global recommendation systems (Bhagavatula et al. Our proposed task combines quotability identification, iden- 2018; Strohman, Croft, and Jensen 2007; He et al. 2010; tifying and extracting quotable phrases from a source docu- Gupta and Varma 2017) use the entire text of the new doc-
ument to recommend candidate documents to cite. Our pro- Though the methodological solutions for QA and our pro- posed task is most similar to local CR, where models con- posed QR task are identical, there are differences in how end dition their recommendations on only a small window of to- users would use model predictions. In QA, for a given ques- kens near the value to be predicted rather than the entire new tion and set of documents, there are known paragraph(s) in document (Lu et al. 2011; Huang et al. 2015; He et al. 2010, the documents which contain the correct answer(s). Rank- 2011; Peng, Liu, and Lin 2016; Fetahu et al. 2016; Jeong ing the paragraphs is only an intermediate step that assists et al. 2019). However, unlike CR, QR models must learn to in finding the correct answer span. End users are expected make finer-grained recommendations, identifying both para- to only read the predicted answer, not the entire paragraph. graphs and arbitrary length subspans in a source document However, as described in the findings of our pilot study that might be most relevant and quotable. Thus, although ex- and crowdsource evaluation, in QR there are often multiple isting CR systems could be applied to the paragraph ranking paragraphs and spans an author could reasonably choose in portion of our QR task by treating each source paragraph as a given context. Retrieving topically relevant and quotable a different document, they would not be able to also identify paragraphs is therefore an equally important task, as authors and recommend specific quotable spans of tokens in those may read multiple top ranked paragraphs and their suggested paragraphs. spans as they search for a quote. Most prior CR work has been applied to the domain of scientific articles, recommending citations for an academic Quote Recommendation paper, but applications to news data, predicting which news At a high level, QR is a content-based recommendation articles will be cited in a given context, have also been ex- task – recommending relevant and quotable content from a plored (Fetahu et al. 2016; Peng, Liu, and Lin 2016). Re- source document to an author as they compose a new docu- cently, neural networks have shown strong performance on ment. Specifically, given a snippet of context in a new doc- citation and other content-based recommendation tasks, out- ument and the text of a single source document that the au- performing traditional bag-of-words and citation translation thor wants to quote, we learn both to rank the paragraphs models (Huang et al. 2015; Bhagavatula et al. 2018; Ebesu of the source document by their relevance and quotability and Fang 2017; Zhang et al. 2018; Jeong et al. 2019). Cur- and to extract quotable spans of text from those paragraphs. rently, BERT-based models achieve SoTA performance on To give the most relevant recommendations, this snippet of local recommendation problems (Jeong et al. 2019). Though context should contain the content an author has written up the BERT-based model of Jeong et al. (2019) uses additional to the point they are recommended and select a quote. Thus, citation network features to make predictions, making it in- we include in each context the sentences in the new docu- compatible with our context-only QR task and unable to ment that precede the quote. We henceforth refer to this as generalize to new source documents not seen during train- the left context. We do not include the part of the sentence ing, the strong performance of BERT further motivates our containing the quote or content after, since those often con- model selection for QR. tain analysis of the quote’s content. Table 1 shows a portion of an example left context. The bold portion of the source Question Answering As described earlier, we cast our paragraph is then quoted in the next sentence of the new proposed QR task as a variant of open-QA. For a given document. We also include the new document’s title as fur- question, open-QA systems first retrieve d relevant doc- ther context as it is acts as compact summary and allows uments, usually with traditional IR models such TF-IDF us to make context-based predictions when there is no left (Chen et al. 2017; Lee et al. 2018; Clark and Gardener context (i.e. the quote occurs in the first sentence). We re- 2018) or BM25 (Wang et al. 2018a,b, 2019). To decrease fer to the concatenation of the new document’s title with the the search space for the machine reader, many systems train quote’s left context in the document as the quote query. a paragraph or sentence-level ranker to estimate the likeli- As described in Related Work section, our proposed QR hood that a passage contains the answer (Clark and Gardener task differs from previously proposed QR task variants (Nic- 2018; Htut, Bowman, and Cho 2018; Lee et al. 2018; Wang ulae et al. 2015; Tan, Wan, and Xiao 2015, 2016; Lee et al. et al. 2018a,b, 2019). Finally, a machine reader is trained 2016; Ahn et al. 2016; Tan et al. 2018). For our task, sys- to score answer spans in the top p ranked paragraphs. At tems must both rank the candidate source paragraphs and prediction time, the reader processes and predicts a span learn to identify specific quotable spans of text in them. Prior in each of the p paragraphs independently, and the overall context-based QR task variants do not include this span pre- confidence of each predicted span is calculated as a com- diction component. Instead, they cast QR as a simply rank- bination of the scores from the document ranker, paragraph ing a predefined list of quotes, proverbs and maxims (Tan, ranker and reader. Recently, BERT (Devlin et al. 2019) has Wan, and Xiao 2015, 2016; Lee et al. 2016; Ahn et al. 2016; shown strong performance on both open-QA (Liu et al. Tan et al. 2018). 2019; Wang et al. 2019) and single paragraph QA datasets, such as SQuAD (Rajpurkar et al. 2016), where no document Pilot Study: Expert Performance or paragraph ranking is required (Alberti et al. 2019). For our To assess the difficulty of our proposed QR task, we conduct QR task, we adapt SoTA BERT-based methods from open- a small pilot study. We recruit two experienced journalists QA (Wang et al. 2019), combining the scores of two sep- and ask them to perform the following two tasks: arately trained BERT models, a paragraph ranker and span predictor, to rank paragraphs and spans. 1. Paragraph selection: Given a context (title & entire left
Paragraph Span Document Title: White House concussion summit opens with Journalist Accuracy EM BOW-F1 introduction from Huntingtown’s Tori Bellucci A 66.6 20.0 56.2 Document Context: ... In his brief remarks Thursday, Obama B 46.7 00.0 46.9 discussed the importance of sports in American society, and the need to ensure the protection of youth who play them. “We want Table 2: Expert Analysis Results: predictive accuracy of the our kids playing sports,” Obama said. two journalists on a sample of 15 quotes from our test set. Predicted Quote: I’d be much more troubled if young people were shying away from sports. As parents, though, we want to keep them safe, and that means we have to have better infor- mation. context) in a new document and all the paragraphs in the source document, identify all of the paragraphs that are Table 3: Example of over-prediction from our pilot study, relevant and might be appropriate to quote in the next sen- where the predicted quote contains both the true quote (bold) tence of the article. and some extra phrases. Over-prediction is a reasonable er- 2. Span selection: Given a context (title & entire left con- ror since authors often paraphrase portions of the surround- text) in a new document and the actual paragraph in the ing content in the introduction to and discussion of a quote. source document that the author quoted in the next sen- tence, predict which text span (can be a sentence, incom- plete sentence, a span of multiple sentences, etc.) was ac- tually quoted. some extra tokens on either side (see example in Table 3). We are interested in examining how difficult it is for experts This was due in part to the journalists’ tendencies to predict to find and suggest relevant and quotable source passages, complete sentences as quotes (70% of predictions). How- and if the expert suggestions line up with the actual passages ever, examining the original news articles, we find that the quoted in the original articles. original authors often paraphrase content from the surround- For our study, we sample from the test split of our dataset. ing paragraph when introducing a direct quote. Thus, over- As detailed in Dataset section, the source documents in prediction is a reasonable error since authors may simply this dataset are Presidential Speeches and the new docu- paraphrase portions of the predicted quote. ments are news articles which report on and quote from the speeches. We sample 1 speech, “Remarks by the President For feedback, the journalists noted that they only closely at the Healthy Kids and Safe Sports Concussion Summit,” considered paragraphs and spans with specific anecdotes and and sample 1 quote from each of the 15 associated news ar- figures that built on the previous context. Both noted that, ticles for inclusion in our analysis. The journalists perform while neither task was difficult, task 1 was harder since they task 1 first, marking all quotable paragraphs for each of the had to search the entire speech and read multiple paragraphs. 15 contexts. Then, using the same contexts, they perform However, one journalist noted that task 1’s search to identify task 2, this time also viewing the positive paragraph quoted relevant paragraphs became easier and faster as they went by the original author. Finally, the journalists were asked to along because they started to memorize the speech. assess each task’s difficulty and list the factors that influ- These findings indicate the usefulness of our proposed QR enced their predictions, focusing on what factors made spe- system as a writing assistance tool – ranking paragraphs by cific paragraphs and spans relevant and quotable. their predicted quotability would help authors find quotes The selected speech contains 28 paragraphs, of which 7 more quickly. This would help prevent authors from having are quoted across the 15 articles. 3 of the 15 quotes are the to nearly memorize each source document to ensure fast re- same excerpt of text – the other 12 are unique, though some call, which might be difficult when authors are also focused overlap. For each quote in task 1, we calculate paragraph on composing the new document or if the source document accuracy as 1 if the journalist identified the positive para- is long (the median speech in our dataset is 43 paragraphs, graph in their list of relevant paragraphs, otherwise 0. Since 1.5x the length of the Concussion Summit speech). the setup for our task 2 is identical to QA, we use 2 com- mon QA metrics: exact match (EM), and macro-averaged Further, the difficulty for experts to identify the para- bag-of-words F1. EM measures if the predicted span exactly graphs and spans quoted by the original authors, along with matches the positive quote, and BOW-F1 measures their av- the median of 2 quotable paragraphs listed per news arti- erage word overlap. cle context, point towards the subjectivity of QR and its dif- As seen in Table 2, identifying the originally quoted para- ferences from QA. In traditional QA datasets, there is only graphs and spans is difficult, even for experts. On task 1, a single correct answer for a given question. On the other journalist A identified the positive paragraph in 10 of 15 in- hand, in QR, multiple paragraphs and quotes could be suit- stances and journalist B in 7. A and B listed a median of 2 able for a given context. Traditional recommendation and (mean 3.8) relevant and quotable paragraphs per news arti- QA metrics, therefore, may be especially harsh for evaluat- cle context. On task 2, both journalists struggled to identify ing the quality of a model’s predictions. Motivated by this the exact quote used by the original author – A identified finding, in later section we perform a crowsourced evalu- 3 of the exact quotes and B none. The most common er- ation of our best model’s predictions to assess whether its ror was over-prediction (74% of errors), where the journal- “incorrect” recommendations (i.e. of passages not selected ist predicted quotes that contained both the true quote and by the original authors) are still relevant and appropriate.
Methods Paragraph Ranking To generate training and testing ex- amples for the models, we iterate over each quote query As discussed in introduction, our QR problem can be viewed and create a (quote query, source paragraph) example pair as a variant of open-QA, where the questions are the quote for each paragraph in the corresponding source document. queries, the passages containing the answers are the differ- Each pair has a binary label: 1 if the author actually quoted ent paragraphs of the source document, and the answers are from the paragraph in the next sentence of the document, 0 the exact quotes the author used in the next sentence of the otherwise. Following Wang et al. (2019), we fine-tune para- document. Thus, to model the QR problem, we apply SoTA graph BERT using the popular softmax-based listwise rank- BERT-based methods from open-QA. As described in Re- ing loss. Listwise loss functions are generally more effec- lated Work section, SoTA open-QA methods consist of three tive at ranking tasks than pointwise or pairwise loss func- parts – 1) a bag-of-words retrieval system, such as BM25, tions (Guo et al. 2019) and have been used successfully with to retrieve an initial set of candidate passages 2) a passage neural ranking architectures (Huang et al. 2013; Shen et al. reranker for selecting high quality passages 3) a machine 2014). reading module which predicts an answer span in a given Thus, a single training example for paragraph BERT con- passage. The final score for a passage and its predicted span sists of n + 1 (quote query, source paragraph) input pairs – is computed as some combination of the scores from the one for the positive-labeled paragraph that contains the real three steps. Since, for our task, we are only focusing on a quote the author chose and n negative, unquoted, paragraphs single source document, the initial retrieval step is skipped from the same source document. Each of the n + 1 packed and we only model the paragraph ranking and span predic- input sequences is fed to BERT independently. We use the tion tasks. final hidden vector C ∈ Rh corresponding to the first input Following Wang et al. (2019), we select BERT as both our token [CLS] as the representation for each of the n + 1 se- passage ranker and span predictor. In addition to achieving quences, where h is the size of the final hidden layer. We SoTA performance on open and single-paragraph QA (De- introduce one task-specific parameter, V ∈ Rh , whose dot vlin et al. 2019; Liu et al. 2019; Yang et al. 2019; Wang et al. product with C is the score for each choice. We transform 2019), BERT has also achieved SoTA performance on pas- the n + 1 scores into probabilities using the softmax func- sage and document retrieval (Dai and Callan 2019; Akkaly- tion: V ·Ci oncu Yilmaz et al. 2019; Nogueira et al. 2019). Since, unlike Pi = P e eV ·Cp QA, users of a QR system are likely to examine and use the p∈P full text of multiple top ranked paragraphs, paragraph re- where P is the set of n + 1 sequences. The loss is the negative trieval performance is more important than just as a means log-likelihood (NLL) of the positive sequence. At inference of achieving better span prediction. time, for each quote query we create a (quote query, para- graph) packed input sequence for each paragraph in the cor- BERT Model Inputs The paragraph and span-level BERT responding single source document and rank them by their models receive the same input and are identically structured predicted scores. except for the final layers, loss functions and example labels In order to select the n negative pairs for training, we sam- (paragraph or span-level labels). Due to memory constraints, ple paragraphs from the same source document as our posi- BERT cannot operate on an entire source document as a sin- tive, quoted paragraph. We explore three negative sampling gle input, so we operate on the paragraph level, with BERT methods – uniform sampling, upweighting paragraphs with processing each (quote query, source paragraph) pair as a high TF-IDF cosine similarity to the quote query, and up- single packed input. weighting paragraphs positionally close to the positive para- graph in the source. For input to BERT we tokenize titles, left contexts and source document paragraphs into WordPieces (Wu et al. Span Prediction The span-level BERT model receives the 2016) and cap them at predetermined lengths chosen as hy- same (quote query, paragraph) packed input sequence as perparameters. Left contexts are capped from the right, so the paragraph ranker. BERT processes and predicts spans they always include the words directly preceding the quot- in each paragraph independently. We use the final hidden ing sentence. We treat the quote query as sequence A and vector Ti ∈ Rh as the representation for each WordPiece candidate source paragraph as sequence B. We add a special in a given paragraph. We follow the standard approach of new token ‘[body start]’ to separate the title from the left casting span prediction as two classification tasks, separately context to give the model a notion of which part of the new predicting the start and end of the span (Devlin et al. 2019; document it is reading (Alberti et al. 2019). Thus the Word- Wang et al. 2019; Chen et al. 2017). Piece input to BERT is: We explore two different approaches for training span- [CLS] Title [body start] Context [SEP] Paragraph [SEP] level BERT – positive-only (Chen et al. 2017) and shared- By packing two sequences into the same input, BERT can normalization (Clark and Gardener 2018; Liu et al. 2019; use self attention to model interactions between the se- Wang et al. 2019). In the positive-only approach, span-level quences across multiple levels (word, sentence, paragraph). BERT is trained with only (quote query, paragraph) pairs Since there may be many topically related paragraphs in a that contain the quote the author selected. We discard all source document with differing amounts of quotable con- negative pairs from the training set, and span-level BERT tent, a model must also learn to judge how interesting and is only trained on one pair at a time. We introduce start and meaningful specific content in a paragraph is. end vectors, S, E ∈ Rh . The probability of word i being the
start of the quoted span is the dot product S ·Ti followed by a softmax over all WordPieces in the example. We follow the same approach for calculating the probability of word i be- ing the end of the span using E. The loss is calculated as the average NLL of the correct start and end positions, i.e. the tokens in the paragraph the author actually quoted. Follow- ing Devlin et al. (2019), at prediction time, the score of the span from position i to j is S ·Ti +E ·Tj . We consider a span valid if j > i and i and j occur in the paragraph portion of the input. We retain a mapping from WordPieces to original tokens for prediction. Just as Chen et al. (2017), we adapt the positive-only model to the multi-paragraph setting at in- ference by running it on each paragraph in the correspond- ing source document independently, finding the maximum Figure 1: Combined BERT model architecture. scoring span in each paragraph, then using the unnormalized score (before the softmax) to rank the spans. As first noted by Clark and Gardener (2018), applying this approach to the multi-paragraph setting may be problematic. As the model is trained on and predicts spans in paragraphs independently, pre-softmax scores across paragraphs are not guaranteed to be comparable. Further, since the model is only trained on positive paragraphs that contain true quote spans, it may not effectively learn to score spans in nega- tive paragraphs. To resolve this, we utilize the most effec- tive approach found by Clark and Gardener (2018), shared- normalization, which has also achieved SoTA performance with BERT-based QA systems (Liu et al. 2019; Wang et al. 2019). Just as for the paragraph model, we create n + 1 in- put sequences (1 positive, n negatives), and each input is Figure 2: Distribution over # paragraphs / transcript. processed independently by span-level BERT. We then use a modified objective function where the softmax operation is shared across all positions in the n + 1 pairs to calculate the the paragraph and span model scores using late fusion (Fig- start and end probabilities. Thus the probability of word j in ure 1). We calculate the probability p(pi |qj ) of paragraph paragraph i starting the span is: pi given quote query qj as the softmax over the scores of eS·Tij Pij = P P eS·Tpw all possible source paragraphs for quote query j. The prob- p∈P w∈p ability p(sik |pi , qj ) of span sik , the maximum scoring span where w iterates over each word in each of the n + 1 ex- in paragraph i, given quote query qj and its paragraph pi is amples. This forces the model to output comparable scores calculated as the softmax over the maximum scoring span in across multiple paragraphs and trains it to learn patterns ef- each of the possible source paragraphs for qj . Following Lee fective in scoring both positive and negative paragraphs. The et al. (2018), we calculate the overall confidence of a para- loss is the average NLL of the correct start and end positions graph pi and its predicted answer span sik given the quote in the positive paragraph. At prediction, the model is run on query qj as each paragraph in the corresponding source document inde- score(sik , pi |qj ) = p(sik |pi , qj )α p(pi |qj )β pendently, and the unnormalized scores are used to rank the Each score’s importance is determined by the hyperparame- top span in each paragraph. ters α, β ∈ {0, 0.5, ..., 10}, which we tune on our dev set. Alternatively, we could train a single model to jointly Combined QR Model Due to the success in combining score paragraphs and spans. We tried training separate task- the scores of paragraph and span models in open-QA (Wang specific parameters for paragraph ranking and span predic- et al. 2018a; Lee et al. 2018; Wang et al. 2019; Htut, Bow- tion on top of a shared BERT model, then averaging the man, and Cho 2018), we explore a similar combination of paragraph and span losses. However, this joint model has the paragraph and span-level BERTs for our task. We hy- worse performance on our dev set, so we focus only on the pothesize that the paragraph and span models learn to iden- combined, separately-trained model for our experiments. tify complementary aspects of quotable passages, so com- bining the scores will help rank paragraphs such as those which, although topically similar to the left context, con- Dataset tain differing amounts of quotable language, or paragraphs We use the QUOTUS dataset for our experiments (Nicu- that contain very quotable phrases but are unrelated to the lae et al. 2015). The dataset consists of two sets of texts left context. Thus we create a single, aggregate score for a – transcripts of US Presidential speeches and press confer- paragraph and its top ranked span at inference by combining ences from 2009-2014 and news articles that report on the
Figure 3: Distribution over # tokens / transcript paragraph. Figure 4: Distribution over fraction of paragraph quoted. speeches. The authors identify the quotes in each news ar- ticle that originate from the transcripts. They release the transcripts and the collection of aligned quotes, containing the text of the quote in the news article, its aligned position within the source transcript, and the corresponding news ar- ticle metadata (title, url, timestamp).1 We crawled the provided news article URLs on Mar 15, 2019 and were able to download 110,023 news articles (≈ 75% of the provided URLs). We extracted the title and body content of each article using BeautifulSoup2 and re- moved duplicates, articles with less than 100 characters of body content, and articles where we could not locate the quotes. This leaves our final dataset of 131,507 quotes from Figure 5: Distribution over quote length (# tokens). 3,806 different White House transcripts across 66,364 news articles. The quotes are grouped into 58,040 clusters, groups of quotes, possibly with minor edits, which all align uniformly distributed over the first 80% of a given paragraph to roughly the same place in the transcripts. We then pre- and end positions over the last 80%. processed articles and transcripts with CoreNLP’s sentence We split our dataset into train, dev and test by transcript segmenter and word tokenizer (Manning et al. 2014). publication date. We select transcripts from 2009-01-01 to We use the alignments between the transcripts and news 2012-11-07 as our training set, 2012-11-08 to 2013-06-19 articles to create the labels for our dataset, learning to rec- as our dev set, and 2013-06-20 to 2014-12-31 as our test ommend paragraphs and spans from a source transcript to set (see Table 4). The split dates were selected to ensure no a journalist as they write a news article. For each aligned news articles quote from transcripts in different data splits. quote, we construct the quote query as news article title + left Nearly all (93%) of the news articles in our dataset con- context leading up to the specific quote. The positive span tain only quotes which all align to the same source tran- is the exact span of tokens in the transcript that are aligned script. This is reasonable as a journalist is much more to the quote in the article. The positive paragraph(s) are likely to quote from the most recent and topically relevant those transcript paragraph(s) that contain the positive span source speech than search over less relevant and up-to-date (≈ 99.5% of positive spans are contained in one single para- speeches (Niculae et al. 2015). graph and do not cross paragraph boundaries). The average For the 0.5% of quotes in the training set that span multi- transcript contains 71 paragraphs, (median 43, Figure 2), and ple paragraphs (≈ 500 quotes), we create separate span and the average paragraph is 49 tokens long (median 45, Fig- paragraph training examples for each part, independently ure 3). Similar to findings in studies of media coverage of sampling negative paragraphs for each. For evaluation, we political debates (Tan, Peng, and Smith 2018), paragraphs positively label all p spanned paragraphs, concatenate the earlier in the speech tended to be quoted more often than spans across the p positive paragraphs as the positive span, those later. A positive span, on average, covers 32% of the and concatenate the predicted spans from the top p ranked corresponding positive paragraph (median 24%, Figure 4) and is 21 tokens long (median 17, Figure 5). Only 3% of quotes span an entire paragraph. Many positive spans either # Quote # Speech # Article start at the beginning (13%) or end at end (32%) of the posi- Train 94,931 2,823 47,698 tive paragraph. Otherwise, span start positions are relatively Dev 15,472 391 7,763 Test 21,104 592 10,903 1 http://snap.stanford.edu/quotus/#data 2 https://www.crummy.com/software/BeautifulSoup/bs4/doc/ Table 4: Preprocessed dataset summary statistics.
paragraphs to create the predicted span. Since, as noted in the pilot study, there are often multiple relevant paragraphs and quotes for a given context, evaluat- Model Experimental Settings ing models in these single-answer settings is overly strict. Thus, in the next section, we perform a large-scale crowd- Parameter Settings We cap titles, left contexts and source source evaluation of our best model’s predictions, assessing document paragraphs at 20, 100 and 200 WordPieces, re- whether its “incorrect” recommendations are still relevant spectively. We further explore the effects of decreased article and appropriate. context on predictive accuracy. We fine-tune the publicly re- leased BERT Base3 for our paragraph and span models and Paragraph Ranking Baselines We compare our para- perform a search over batch size ∈ {16, 32} and learning graph ranker against 2 common bag-of-words retrieval mod- rate (Adam) ∈ {5e − 5, 3e − 5, 2e − 5}. We train our mod- els – TF-IDF cosine similarity and BM25 (Jones et al. els for a maximum of 4 epochs and select hyperparameters 2000). We also evaluate performance of Doc2Vec cosine and perform early stopping using our dev set. We also per- similarity (Le and Mikolov 2014) to see if it can capture form a search over n ∈ {3, 6, 9, 12} sampled negative para- semantic similarity between the left context and paragraph. graphs per positive paragraph for our paragraph and shared- Finally, we evaluate two neural ranking baselines – K-NRM norm models. In preliminary experiments on our dev set, (Xiong et al. 2017), a popular kernel pooling neural network we find no significant differences between uniform random that uses a translation matrix of word embedding similar- sampling and our proposed non-uniform sampling schemes. ities between a query and document to learn-to-rank doc- Instead, model performance generally increases by sampling uments for ad-hoc retrieval, and Conv-KNRM (Dai et al. a larger number of negatives. We suspect this is because a 2018), which builds on K-NRM by using CNNs to model larger sample will likely include both hard- and easy-to-rank soft n-gram matches between a query and document. negatives, thus training the model to differentiate between As noted in Related Work, previous approaches to QR the positive paragraph and both semantically similar and dis- use different tasks setup (e.g. ranking a predefined list of similar paragraphs. The best paragraph model is trained with quotes), and the proposed models use additional, non-textual 12 negative examples, batch size 32, and learning rate 2e-5. features, such as popularity data crawled from the web (Tan, The best shared-norm span model is trained with 9 negative Wan, and Xiao 2015, 2016). Similarly, the SoTA citation examples, batch size 16, and learning rate 3e-5. The best recommendation system (Jeong et al. 2019) makes use of ci- positive-only span model is trained with batch size 32 and tation network metadata features in addition to BERT. Due learning rate 3e-5. to these incompatibilities with our QR task, where models are only provided with the texts of the context and source, Metrics To evaluate paragraph rankings, we use mean we do not evaluate these previous systems. average precision (mAP) and top-k accuracy, defined as For each baseline, we calculate the similarity between a the proportion of examples where the positive paragraph is quote query and each possible paragraph in the correspond- ranked in the top k. Since only ≈ 100 quotes in the test set ing source document, just as for our BERT models. For span multiple positive paragraphs, mAP is nearly equivalent Doc2Vec we use the distributed bag-of-words algorithm and to mean reciprocal rank. For top-k accuracy, we require that search over the hyperparameters suggested by Lau and Bald- at least 1 of the positive paragraphs is ranked in the top k. win (2016). For K-NRM and Conv-KNRM, we use the sug- Since the setup for our span prediction task is identical to gested hyperparameters from Xiong et al. (2017) and Dai QA, we evaluate the span-level models using the same two et al. (2018) and use pre-trained Glove embeddings (Pen- QA metrics that we used to evaluate the expert predictions nington, Socher, and Manning 2014). We train the models in the pilot study – exact match (EM), and macro-averaged on our training set and use early stopping on our dev set. bag-of-words F1. EM measures if the predicted span exactly For each baseline we search over the number of tokens of matches the positive quote, and BOW-F1 measures their av- left context ∈ {10, 20, 30, ..., 100} to include in the article erage word overlap. We evaluate the span models in two set- query in addition to the title, and find 30 tokens works best tings, positive and top. In the positive setting, we assume that for Conv-KNRM and 40 for all other baselines on the dev an author has already selected the positive paragraph and set. is interested in the recommended span for that paragraph. Thus, for a given test example, we only evaluate our model’s Results & Analysis predictions on the positive, quoted paragraph(s) against the true quote. This is equivalent to a closed, single paragraph Paragraph Ranking QA evaluation. In the top setting, for a given quote, we run Table 5 shows the paragraph ranking results for all the meth- the model on all possible paragraphs in the corresponding ods with their best settings. We test significance with a per- source, identify the top scoring span in each paragraph, rank mutation test, p < 0.01. We find that paragraph BERT (#6) the spans by their scores, and evaluate the top ranked span. significantly outperforms all baselines (#1-5) by a large mar- This setting is equivalent to open-QA evaluation. Though gin, outperforming BM25, the best baseline by ≈ 13% ab- more realistic, this is harder, as the model may not rank the solute and 30 − 40% relative across all metrics. Examining predicted span in the positive paragraph the highest. the baselines, we find that Doc2Vec is much worse than the other baselines that can model exact token matches between 3 passages. We hypothesize that these exact matches, particu- https://storage.googleapis.com/bert models/2018 10 18/ uncased L-12 H-768 A-12.zip larly of named entities, are a strong indicator of transcript
Method mAP Acc@1 Acc@3 Acc@5 1. TF-IDF 39.0 27.3 42.5 50.7 2. BM25 39.9 28.4 43.5 51.5 3. Doc2Vec 31.7 21.1 33.4 41.4 4. K-NRM 39.2 27.3 42.6 51.1 5. Conv-KNRM 38.9 27.2 41.9 50.4 6. Paragraph 52.4 39.7 58.1 67.3 7. Positive-only Span 32.2 20.0 34.3 44.2 8. Shared-norm Span 51.3 38.6 56.8 65.9 9. Combined (6 + 7) 52.7 40.0 58.1 67.7 10. Combined (6 + 8) 53.2 40.5 59.1 68.2 Table 5: Paragraph ranking results. Bold entries significantly outperform all other models (p < 0.01). Models 6-10 are Figure 6: Distribution over absolute distance (in number of BERT models. Acc is short for top-k accuracy. paragraphs) between the top paragraph incorrectly recom- mended by the combined BERT model and the positive para- graph chosen by the original journalist. paragraph relevance. For instance, if the journalist is dis- cussing health care and mentions the Affordable Care Act in the left context, then paragraphs that not only are topi- forming all other models across all ranking metrics. It im- cally related to health care, but also contain that specific en- proves over the single paragraph BERT model by ≈ 1% ab- tity may be more relevant. BERT’s ability to model both ex- solute across each metric and by ≈ 15% over BM25, the best act and semantically similar matches between a context and baseline. This combined model upweights the paragraph paragraph across multiple levels of granularity might lead score relative to the shared-norm score: α = 3, β = 9.5. To to its improved performance over the token and embedding evaluate if ranking performance can be boosted even while matching baselines. adding a weaker span ranker, we also experiment with com- However, lexical and semantic similarity to the context bining the paragraph model and positive-only span model. is only one aspect of a paragraph’s quotability (Danescu- The best performing combined positive-only model (#9) up- Niculescu-Mizil et al. 2012). In order to be selected by a weights the paragraph score α = 1.5, β = 7.0. Though journalist, a paragraph must not only be topically relevant, it does not perform as well as the combined shared-norm but also contain material that is interesting, meaningful, and model, the combined positive-only model improves over serves the journalist’s purpose. All of our baselines rank the single paragraph model in mAP and top 1 and 5 ac- paragraphs solely by similarity to the context and fail to curacy. This further validates our hypothesis that the span model the quotability of the paragraph. BERT’s superior per- and paragraph-level models learn complimentary aspects of formance in scoring hard-to-rank paragraphs, such as those quotable paragraphs. which all discuss the same event, person, or topic as the left context, but contain varying amounts of quotable language, Error Analysis To better understand the errors our best indicates that it may also capture aspects of quotability be- combined BERT model (#10) makes, we examine the 12,557 yond relevance. quotes in the test set where it does not rank the positive As the span-level BERT models can also be used to rank paragraph as the top 1. We focus our analysis on the top- paragraphs, we evaluate the positive-only (#7) and shared- 1 predicted (incorrect) paragraph, and measure the absolute norm (#8) span models on the paragraph ranking task. For distance (in # of paragraphs) between it and the positive, a given quote, we run the span models on all possible para- actually quoted paragraph. As can be seen in figure 6, we graphs from the corresponding source and rank each para- find that our model often recommends a paragraph nearby graph by its maximum scoring span. Though training with the one selected by the journalist. For approximately 20% shared-norm loss substantially improves ranking perfor- of quotes, the model recommends a paragraph directly adja- mance over the positive-only model, paragraph BERT (#6) cent to the positive paragraph, and for roughly 38%, it rec- significantly outperforms both span BERT models across ommends a paragraph within 3 positions of the positive. As all metrics. However, although paragraph BERT is over- adjacent paragraphs in a speech are more likely to be on the all a stronger ranker, it does not outperform shared-norm same topic, this confirms that our model’s incorrect recom- BERT on every test example. On 175 examples, shared- mendations are often still topically similar to the paragraph norm BERT ranks the ground truth as the top 1, but para- the journalist chose. graph BERT ranks it outside the top 5. We hypothesize that this is due to differences in the granularities of their objec- Span Prediction tives, with each model identifying different aspects of what As seen in Table 6, span prediction is harder than paragraph makes a paragraph quotable and relevant. ranking. As baselines, we use the entire paragraph (#1) or its Finally, our experiments combining the paragraph and first (#2) or last (#3) sentence as the predicted span. These span models parallel findings in open-QA (Wang et al. paragraphs are either the positive paragraph (positive set- 2019), with the combined paragraph and shared-norm model ting) or the top ranked paragraph (top setting) by the best yielding the best performance (#10) and significantly outper- single paragraph model (ref. Table 5 #6). The positive-only
EM EM F1 F1 quoted phrase is still relevant since the journalist describes Method Positive Top Positive Top the team’s successes in the context introducing the quote. 1. Paragraph (Full) 2.9 0.6 42.2 24.4 2. Paragraph (1st-Sent) 4.8 1.2 28.3 10.7 Crowdsourced Evaluation 3. Paragraph (Last-Sent) 5.5 1.4 30.6 12.8 4. Positive-only Span 21.0 7.8 57.4 23.6 As noted in the findings of our pilot study and error analysis, 5. Shared-norm Span 22.8 12.0 59.0 34.1 there are often multiple quotes and paragraphs a journalist 6. Combined (1 + 4) 21.0 11.2 57.4 33.8 could select for a given context. Thus we hypothesize that 7. Combined (1 + 5) 22.8 12.0 59.0 34.5 our model’s “incorrect” top recommendations (non-positive quotes) could still be relevant and appropriate. Due to the Table 6: Span prediction results of BERT models. Bold en- limited time of the two journalists, we turn to crowdworkers tries significantly outperform all other models. to perform a large-scale evaluation our model’s predictions. We focus this evaluation on paragraph rather than span predictions since we believe authors writing a new docu- span model (#4) significantly outperforms the baselines in ment will use the entire paragraph to 1) better understand the EM positive & top and F1 positive, but its predictions are broader source context and 2) once a specific quote is cho- worse than the full paragraph on F1 top. Parallel to findings sen, paraphrase other content from the paragraph to intro- in open-QA (Clark and Gardener 2018; Wang et al. 2019), duce and contextualize it. We randomly sample 1,000 quotes the shared-norm model (#5) significantly outperforms the from our test set where our best combined model does not positive-only model and baselines across all metrics, even rank the positive paragraph as the top 1 and evaluate, using EM and F1 positive, demonstrating that the shared loss func- Mechanical Turk crowdworkers, whether the model’s top tion forces the span model to produce more comparable ranked (non-positive) paragraph is relevant and quotable. scores between different paragraphs and improves its abil- For each quote, crowdworkers see the article title, left ity to identify spans in positive paragraphs. context (100 WordPieces, converted to regular tokens and Just as with findings in open-QA and our paragraph rank- rounded up to a full sentence), and the top non-positive para- ing task, combining the paragraph and shared-norm mod- graph ranked by our best model. They are asked “Does this els improves performance (#7). Note, the performance in paragraph contain an excerpt that would be appropriate to the positive setting is identical as we only compute over- quote in the next sentence of the news article?” We block all confidence scores for the top ranked span in each para- workers who do not follow instructions and collect 5 judg- graph. The paragraph score will not influence the selection ments per question. Each worker labeled 10 questions, for of a span in a given paragraph, only the ranking of those a total of 500 unique workers. Inter-annotator agreement is spans across paragraphs. The differences between the best 0.26 (Fleiss’ κ), indicating fair agreement. We aggregate la- combined shared-norm model (#7) and the single shared- bels using majority vote. norm model (#5) are not as pronounced as in the ranking Workers mark our predicted paragraph as relevant 85% of task. The combined shared-norm model only significantly the time, suggesting that our model’s top predictions seem outperforms the single shared-norm model on F1 top by 1% quotable and meaningful to non-experts even if they are relative, with no improvement on EM top. We also evaluate not the paragraph chosen by the original journalist. Table 7 the combined positive-only model (#6). Though worse than shows an example context and top ranked paragraph, along both shared-norm models, it improves ≈ 40% relative over with the true next sentence in the original article. The para- the single positive-only model in EM and F1 top. graph is marked as quotable by all 5 workers, and, in fact, the selected quote in the original article is similar to the last Error Analysis To further assess the combined shared- sentence of the top ranked paragraph. norm model (#7), we examine its errors on the test set, com- paring its predictions on the positive paragraph to the true Ablation Studies quote. In some instances, we find that the model recom- To explore the effects of decreasing the amount of article mends a span that is semantically similar to the true quote content, we perform 2 ablation experiments, training new or one that is quoted multiple sentences later. For 15% of er- paragraph and span models and combining their scores using rors, we find that the model under-predicts, predicting a sub- the same hyperparameters as our best models. For brevity, phrase in the true quote. However, parallel to the pilot study Table 8 shows results for only a subset of the ranking and findings, the most common error is over-prediction (47% of span prediction metrics. Decreases across all metrics are errors), where the model predicts a span containing the true consistent with the shown subset. First, we remove the ar- quote and some phrases surrounding it. ticle title, but keep 100 WordPieces of left context. Perfor- Further, as in pilot study, we find that in many instances mance (#2) decreases by ≈ 2% and 1% absolute across rank- of over-prediction the author paraphrases the predicted sur- ing and span metrics, respectively. This decrease is expected rounding phrases in the introduction to or discussion of as an article title provides both a high level overview of the the quote. For instance, on the example shown in Table 1, article and the only context when the quote occurs in the be- the model over-predicts “At one point, they won 27 games ginning of the article. However, in domains such as news, straight- the second-longest winning streak ever, extraordi- where the title is often written after the body of the arti- narily impressive- almost as impressive as the Bulls’ 72-win cle, performance on this experiment may be more realistic. season” (true quote italiced). Though incorrect, the non- Next, we keep 20 WordPieces of title content, but vary the
Document Title: Janet Yellen to become the first female US source documents of any kind, such as press releases, meet- Fed boss ing transcripts, and book chapters. Document Context: ... Yellen will chair the Federal Reserve For domains where authors may quote from multiple board, replacing Ben Bernanke after his second four-year term comes to an end at the beginning of next year. source documents with a very large number of total para- Top Ranked Source Paragraph: So, Janet, I thank you for graphs, we could explore how the BERT model scales and taking on this new assignment. And given the urgent economic determine whether we would need to train a document challenges facing our nation, I urge the Senate to confirm Janet ranker to filter documents as a first step. For applications without delay. I am absolutely confident that she will be an ex- where the right context may be partially written before se- ceptional chair of the Federal Reserve. I should add that she’ll lecting a quote, we could explore the effects of including be the first woman to lead the Fed in its 100-year history. And bidirectional context on predictive performance. Finally, to I know a lot of Americans- men and women- thank you for not improve identification of quotable content, we could exper- only your example and your excellence, but also being a role iment with adding manually designed quotability features model for a lot of folks out there. to BERT’s learned representations (e.g. use of positive or True Next Sentence: “American workers and families will have a champion in Janet Yellen,” President Obama said. negative words or personal pronouns) (Danescu-Niculescu- Mizil et al. 2012; Bendersky and Smith 2012; Tan, Peng, and Table 7: A top ranked (non-positive) paragraph that is Smith 2018) or learn them automatically by training a lan- marked as quotable by participants in our crowdsourced guage model on a large collection of quotes, such as Meme- evaluation. The paragraph is topically relevant to the doc- tracker (Leskovec, Backstrom, and Kleinberg 2000). ument context and contains many specific figures and anec- dotes. Its last sentence (italic), just as the ground truth quote, Acknowledgements discusses Yellen’s relationship to everyday Americans. We would like to thank journalists Megan Chan and Ashley Edwards for their participation in our pilot study and help- Paragraph Ranking Span Prediction ful feedback. We also thank Jason Baldridge and the anony- Acc@1 Acc@5 F1 (Pos) F1 (Top) mous reviewers for their helpful and insightful comments. 1. Full Model 40.6 68.8 59.0 34.5 2. No Title - 1.9 - 2.0 - 0.7 - 1.2 References 3. Context: 50 - 1.0 - 1.3 - 0.4 - 0.5 Ahn, Y.; Lee, H.; Jeon, H.; Ha, S.; and goo Lee, S. 2016. 4. Context: 25 - 2.7 - 3.5 - 1.1 - 1.7 Quote Recommendation for Dialogs and Writings. In CBRec- 5. Context: 10 - 8.3 - 9.0 - 2.7 - 5.3 Sys@RecSys. Table 8: Performance of combined model on dev set with Akkalyoncu Yilmaz, Z.; Wang, S.; Yang, W.; Zhang, H.; and Lin, varying amounts of quote query context. The full model is J. 2019. Applying BERT to Document Retrieval with Birch. In EMNLP-IJCNLP: System Demonstrations. our best performing combined model that uses the title and 100 WordPiece tokens as the left context. Acc is short for Alberti, C.; Lee, K.; ; and Collins, M. 2019. A bert baseline for the top-k accuracy. Pos is short for positive. “-” indicates per- natural questions. In arXiv preprint arXiv:1901.08634. formance relative to our best model. Bendersky, M.; and Smith, D. A. 2012. A dictionary of wisdom and wit: Learning to extract quotable phrases. In NAACL Workshop on Computational Linguistics for Literature. left context WordPieces ∈ {50, 25, 10}. Performance across Bhagavatula, C.; Feldman, S.; Power, R.; and Ammar, W. 2018. all metrics decreases with context size, with the sharpest Content-Based Citation Recommendation. In NAACL-HLT. drops from 25 to 10 WordPieces. Encouragingly, for data Chen, D.; Fisch, A.; Weston, J.; and Bordes, A. 2017. Reading and tasks where the left context size is limited, the 50 Word- wikipedia to answer open-domain questions. In ACL. Piece model (#3) only loses ≈1-2% relative performance Clark, C.; and Gardener, M. 2018. Simple and Effective Multi- across each metric. Though these models can use fewer to- Paragraph Reading Comprehension. In ACL. kens, the tokens are at the end of the context right before the Clark, E.; Ross, A. S.; Tan, C.; Ji, Y.; and Smith, N. A. 2018. Cre- quote. We hypothesize that, due to this proximity, they are ative Writing with a Machine in the Loop: Case Studies on Slogans likely to contain strong signals for prediction. and Stories. In 23rd International Conference on Intelligent User Interfaces. Conclusion and Future Work Cole, P. 2008. How to Write: News writing. The In this paper, we introduce a new context-based quote rec- Guardian. URL https://www.theguardian.com/books/2008/sep/25/ ommendation task, using the content an author has already writing.journalism.news. Last accessed: March 25, 2021. written to rank the paragraphs and score the spans of a Dai, Z.; and Callan, J. 2019. Deeper Text Understanding for IR source document by their predicted relevance and quotabil- with Contextual Neural Language Modeling. In SIGIR. ity. We cast the task as a variant of open-domain question an- Dai, Z.; Xiong, C.; Callan, J.; and Liu, Z. 2018. Convolutional swering, and explore applications of state-of-the-art BERT Neural Networks for Soft-Matching N-Grams in Ad-hoc Search. models to the problem. Though we conduct experiments on In WSDM. a dataset of speeches and news articles, our proposed ap- Danescu-Niculescu-Mizil, C.; Cheng, J.; Kleinberg, J.; and Lee, L. proach could be used in many domains. Bloggers, students, 2012. You Had Me at Hello: How Phrasing Affects Memorability. and journalists could use our method to rank passages in In ACL.
You can also read